How we test

This page states exactly what DeAIze measures, how a rewrite is accepted, and how to read the numbers we publish. It is written so that anyone can check our claims against their own text.

1. What a detection score is

Detectors do not detect authorship; they detect statistical regularity. Machine-generated text tends to be more predictable than human writing — lower perplexity, flatter sentence rhythm, fewer lexical surprises. Every score you see, ours included, is a probability estimate from that signal, not a measurement of who wrote the text.

2. Sentence-level scoring

A single document score tells you nothing about where the problem is. Highlight AI traces per sentence — pinpoint exactly which lines to rewrite. The per-sentence scores are what the rewrite loop acts on, and they are what the workbench highlights.

3. The closed loop

Score → per-sentence feedback → rewrite → re-score until it passes. A rewrite candidate is accepted only when semantic similarity to the original clears a threshold, which is why meaning is preserved and why the tool reports an "auto-refund on no-op" when it cannot improve a passage rather than inventing a change.

4. Detector cross-checks

A prediction from our own model is not evidence. Where a detector API is configured, the verify step sends the rewritten text to the real detector and records that result. The count of those runs is published on the statistics page so the number of third-party checks behind our figures is visible.

5. What the "predicted human score" means

It is our own model's estimate of how human-written the output reads — a directional metric for tracking improvement, not an official score from any third-party detector. When we quote a third-party number (GPTZero, Turnitin, Copyleaks, Originality.ai, Sapling) we say so and, where we can, record the run as a cross-check.

6. Limits and disclosure

• Detector scores vary by version, language, text length and domain. A number without a date and a detector version is not reproducible, so we date our published figures.
• Non-English performance is weaker than English, and short texts are noisier than long ones.
• We do not publish accuracy or benchmark figures we cannot reproduce from a stated sample; when we publish a benchmark it carries its sample size, test date and detector versions.
• False positives are real and harmful. If a detector flags your own writing, humanizing your draft is a legitimate response to a probabilistic gatekeeper — misrepresenting someone else's work is not.

Corrections

If any figure here cannot be reproduced, tell us and we will correct it: [email protected].